- Email: [email protected]
Testing causal relationships between wholesale electricity prices and primary energy prices
Energy Policy 62 (2013) 869–877
Contents lists available at ScienceDirect
Energy Policy journal homepage: www.elsevier.com/locate/enpol
Testing causal relationships between wholesale electricity prices and primary energy prices Tadahiro Nakajima a,n, Shigeyuki Hamori b,1 a b
The Kansai Electric Power Company, Incorporated, 3-6-16, Nakanoshima, Kita-Ku, Osaka 530-8270, Japan Faculty of Economics, Kobe University, 2-1, Rokkodai, Nada-Ku, Kobe 657-8501, Japan
H I G H L I G H T S
We test the Granger-causality among wholesale electricity and primary energy prices. We test not only the causality in mean but also the causality in variance. The results show that gas prices Granger-cause electricity prices in mean. We find no Granger-causality in variance among these variables.
art ic l e i nf o
a b s t r a c t
Article history: Received 31 July 2012 Accepted 8 July 2013 Available online 2 August 2013
We apply the lag-augmented vector autoregression technique to test the Granger-causal relationships among wholesale electricity prices, natural gas prices, and crude oil prices. In addition, by adopting a cross-correlation function approach, we test not only the causality in mean but also the causality in variance between the variables. The results of tests using both techniques show that gas prices Grangercause electricity prices in mean. We find no Granger-causality in variance among these variables. & 2013 Elsevier Ltd. All rights reserved.
Keywords: Electricity price Primary energy price Causality test
1. Introduction In the past, it was considered natural that competition was restricted and companies operated monopolistically in the electricity industry. Around 1990, however, the industry began to be liberalized, as many countries deregulated their electricity industries. The wholesale electricity industries in most states of the United States are now open to competition among players in the market. Market structures differ by geographical area, with negotiated- and organized-transaction markets being the two main market types. In a negotiated-transaction market, electricity transactions are directly negotiated by sellers and buyers, and individual transmission line owners establish their own electricity supply plans. On the other hand, in an organized-transaction market, an entity independent of market participants controls all transmission facilities and operates a spot market, with a few exceptions.
n Corresponding author. Tel.: +81 6 6441 8821; fax: +81 6 6441 3584. E-mail addresses: [email protected] (T. Nakajima), [email protected] (S. Hamori). 1 Tel./fax: +81 78 803 6832.
0301-4215/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.enpol.2013.07.033
In December 1999, the United States Federal Energy Regulatory Commission (FERC) issued Order No. 2000 regarding the formation of Regional Transmission Organizations (RTOs). This order asked all the electric utilities owning and operating a transmission system to participate in establishing an organization to operate the system over a large area and in a neutral manner. Entergy TransCo is one such RTO established on the basis of Order No. 2000. This type of RTO is a regulated, for-profit stock company that either owns or leases under long-term contracts all the transmission facilities within a specified area. It is the administrator and operator of the transmission system, responsible for investing in new transmission facilities. The Entergy wholesale market prices are determined at the regional trading hub. An empirical analysis of the relationship between wholesale electricity as a secondary energy source and crude oil, natural gas, coal, and uranium as primary energy sources would contribute greatly to the national energy policy and to the business and trading strategies of those in the energy business. Questions concerning the long-term relationship between the prices of primary energy sources and electricity—such as whether changes in international crude oil prices are useful in predicting regional wholesale electricity price movements or whether there is a volatility spillover effect between the wholesale electricity and
870
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 1 Causality tests of the relationships between electricity prices and primary energy prices in the United States. Sources
Study period Variables electricity Natural gas/Oil
Causality tests between Summary electricity prices and others
Brown and Yücel (2008)
1997–2007
PJM, PV
TZ6-PJM Topock-PV
Woo et al. (2006)
1999–2004
NP15, SP15 (CAISO) PG&E Citygate, SoCal
Emery and Liu (2002)
1996–2000
COB, PV
Henry Hub
PJM
Henry Hub, TZ6, No test among PJM Fuel oil for NY harbor and others
Serletis and Herbert (1999) 1996–1997
Henry Hub, TZ6, Topock
PG&E Citygate-NP15 SoCal-NP15 PG&E Citygate-SP15 SoCal-SP15 Henry Hub-COB Henry Hub-PV
There is bidirectional causality between regional gas and electricity prices, but no causality between gas wellhead and regional electricity prices. There is bidirectional causality between gas and electricity prices.
Electricity prices respond to departure from the equilibrium relationship, but gas prices do not. There appear to be effective arbitraging mechanisms for the prices of gas and fuel oil.
Note: “A-B” means that the past values of “A” give information about the future value of “B.” Serletis and Herbert (1999) argued that the electricity price series is I(0) and each of the other price series is I(1). Therefore, they inferred that the strength of the relationship between the power prices and each of the other prices would be spurious. PJM and PV represent wholesale electricity market prices in the eastern and western United States, respectively. Henry Hub, TZ6, and Topock represent a national natural gas market price, and natural gas market prices in the eastern and western United States, respectively. NP15 and SP15 represent wholesale electricity market prices in northern and southern California, respectively. PG&E Citygate and SoCal represent natural gas market prices in northern and southern California, respectively COB represents a wholesale electricity market price in northern California and Oregon. Fuel oil for NY harbor represents a standard reference price for oil in the U.S. Northeast.
primary energy markets—are of considerable interest from the viewpoint of energy security and energy company management. West Texas Intermediate (WTI) is a light sweet crude oil considered high-quality fuel for power generation. WTI, which is produced in Texas and is delivered at Cushing, Oklahoma, is consumed in the southern states, while the price of WTI is one of the most major benchmarks for the world oil market. Henry Hub is the pricing point for the natural gas in the pipeline system in Louisiana. The price at Henry Hub is seen to be the primary benchmark for the United States gas market. Recently, Henry Hub has started to be seen as the price index for the world natural gas market. Entergy′s service territory includes almost all of Louisiana, the southeastern part of Texas, the eastern three-quarters of Arkansas, and the western half of Mississippi. In other words, Entergy supplies electricity in the region that is a productive center and consuming area of oil and gas, whose prices are international and/or national benchmarks. This region has an extremely rare combination of features. This creates motivation to examine the southern energy market. By testing for the causal relationships among Entergy, WTI, and Henry Hub, we are able to confirm whether WTI and Henry Hub have features of the regional energy market as well as the international and/or national market. It is significant to investigate the causal relationship in the variance between the international and regional energy markets, and between primary and secondary energy markets, because of both electric power suppliers′ and electric power consumers′ concerns. Power suppliers desire to discuss the combination of various power sources, ideal fuel contracts, and the best oil stockpile in order to stabilize the cost of power procurement. Power consumers are interested in the selection and combination of energy sources of their facilities, equipment, and appliances in order to stabilize the total cost of energy procurement when they operate and make investments in these items. We apply the lag-augmented vector autoregression (LA-VAR) technique developed by Toda and Yamamoto (1995) to test for Granger-causality. The LA-VAR approach allows us to test for Granger-causality among time series variables in levels without detecting exactly their integration and cointegration properties.
In this respect, it is more advantageous to use the LA-VAR approach than the vector error correction model (VECM) approaches. The pretests for a unit root and cointegration in the economic time series and the estimation of the cointegrating vector are required before the causality test based on the VECM, and therefore the procedure is very complicated. Moreover, we adopt the cross-correlation function (CCF) approach developed by Cheung and Ng (1996) to test for Granger-causality in volatility as well as level. The traditional tests of causality, developed by Granger (1969) and Engle and Granger (1987), suffer from a number of problems:
The inability to address anything other than mean relationships. Model-building requirements. The need to pay attention to the omission of variables. The CCF approach, in contrast, offers the following advantages. It provides the ability to test for causality not only in the mean but also in the variance, and incorporates the use of univariate model residuals, which makes the building of a multivariate model unnecessary. Many studies empirically analyzed the Granger-causality between power prices and primary energy prices, because most economists can accept the economic model underlying such a Granger-causal relationship. We introduce some studies from the United States that have empirically analyzed the Granger-causality between power prices and primary energy prices using daily data. We summarize these in Table 1. Although researchers have analyzed the relationship between wholesale electricity prices and primary energy prices in the northeastern and western markets of the United States (for example PJM, PV, CAISO, and COB), no studies have examined the southern market (for example the Entergy Hub), to the best of our knowledge. It is difficult to find general or common causality relationships between electricity prices and the primary energy prices, because not many previous analyses exist. However, Brown and Yücel (2008), Woo et al. (2006), and Emery and Liu (2002) argue that past natural gas prices provide information about future wholesale electricity prices. In addition to these studies, Amavilah (1995) and Mohammadi (2009) tested the
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Granger-causality using annual data. Mjelde and Bessler (2009) investigated the Granger-causality using weekly data. Asche, et al. (2006), Serletis and Shahmoradi (2006), Muñoz and Dickey (2009), Nakajima and Hamori (2012), and Nakajima (2013) empirically examined the Granger-causality in the other countries. This paper makes four contributions to the literature. First, we test for a causal relationship between wholesale electricity prices and primary energy prices in southern states. To the best of our knowledge, no similar studies have examined the southern market. Second, we verify the causality hypothesis—which most previous studies propose—that the changes in natural gas prices are useful in predicting power price movements. Third, because few results of causality tests focus on the power market, we accumulate the analytical results. Finally, we analyze the causal relationship between United States wholesale power prices and primary energy prices in the variance for the first time, to our knowledge. The following section explains the LA-VAR technique and the CCF approach. Section 3 describes the data used. Section 4 presents the empirical testing results, and Section 5 summarizes our findings.
2. Methodology We apply the LA-VAR technique developed by Toda and Yamamoto (1995) and the CCF approach developed by Cheung and Ng (1996) in order to test for Granger-causality. The electricity, natural gas, and crude oil prices are probably subject to seasonality. However, it is not clear whether the relationship between these variables has seasonality. To our knowledge, no paper examines seasonality in the Granger-causal relationship between the power prices and other variables. Moreover, because we examine this relationship on the basis of daily data, the seasonality of each variable does not cause a bias in the results of our Granger-causality tests. Therefore, this study does not consider seasonality.
Toda and Yamamoto (1995) proposed a simple technique that requires prior tests of neither integration nor cointegration order. This technique allows the testing of coefficient restrictions in a level VAR model when the variables are of unknown integration or cointegration order. The method proceeds, briefly, as follows. Let the following equation generate fyt g, the n-dimensional vector constituting the level of the variables in the study: yt ¼ g o þ g 1 t þ J 1 yt1 þ J 2 yt2 þ … þ J k ytk þ εt ;
true values of J 1 , J 2 , …, J p are zero, those parameters are not included in restriction (2). Toda and Yamamoto (1995) establish that the Wald statistic asymptotically has a chi-square distribution with degrees of freedom equal to the number of excluded lagged variables regardless of the integration order of the process or the existence of a cointegrating relation. 2.2. The CCF approach Cheung and Ng (1996) developed a two-step procedure to test for causality not only in mean and but also in variance. The procedure is based on the residual CCF. The first step involves the estimation of univariate time-series models that allow for time variation in both conditional means and conditional variances. The second step is conducted by constructing the residuals and squared residuals, both standardized by conditional variances. The CCF of the standardized residuals is used to test the null hypothesis of no causality in mean, and the CCF of squaredstandardized residuals is used to test the null hypothesis of no causality in variance. Suppose two time series can be written as qffiffiffiffiffiffiffi X t ¼ μx;t þ hx;t εt ; ð4Þ Y t ¼ μy;t þ
t ¼ 1; 2; …; T
ð1Þ
where t is the time trend; k is the lag length; g o , g 1 , J 1 , J 2 , …, J k are the vectors or matrices of coefficients; and εt is an i.i.d. sequence of n-dimensional random vectors with zero mean and covariance matrix Σ ε . We formulate the null hypothesis—that the j-th variable does not Granger-cause the i-th variable—as follows: H 0 : J ij ð1Þ ¼ J ij ð2Þ ¼ … ¼ J ij ðkÞ ¼ 0;
yt ¼ γ^ 0 þ γ^ 1 t þ ^J 1 yt1 þ ^J 2 yt2 þ … þ ^J p ytp þ ε^ t
ð3Þ
where circumflex (^) indicates an estimation by OLS and p ¼ k þ dmax represents the true lag length k augmented by a suspected maximum integration order, dmax ðk≥dmax Þ. Considering that the
ð5Þ
Y t μy;t X t μx;t εt ¼ pffiffiffiffiffiffiffi ; ζ t ¼ pffiffiffiffiffiffiffi hy;t hx;t
ð6Þ
2 2 Y t μy;t X t μx;t ¼ ε2t ; vt ¼ ¼ ζ 2t : hx;t hy;t
ð7Þ
As both εt , ζ t and ut , vt are unobservable, we must use their estimates, ε^ t , ζ^ t and u^ t , v^ t to test the hypothesis of no causality in mean and variance, respectively. Next, we compute the sample cross-correlation coefficient at lag k, r^ εζ ðkÞ and r^ uv ðkÞ, from the consistent estimates of the conditional mean and variance of X t and Y t . This leaves us with cεζ ðkÞ r^ εζ ðkÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi cεε ð0Þcζζ ð0Þ
ð8Þ
cuv ðkÞ r^ uv ðkÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi cuu ð0Þcvv ð0Þ
ð9Þ
where cεζ ðkÞ and cuv ðkÞ are the k-th lag sample cross-covariance given by cεζ ðkÞ ¼
1 ^ ∑ð^εt ^ε Þðζ^ tk ζÞ; T
cuv ðkÞ ¼
1 ^ v^ tk v^ Þ; ∑ðu^ t uÞð T
ð2Þ
where J ij ðhÞ is the ði; jÞ element of the matrix J h ðh ¼ 1; 2; …; kÞ. The test can be conducted with the following VAR model, in level form, estimated by ordinary least squares (OLS):
qffiffiffiffiffiffiffi hy;t ζ t ;
where μx;t and μy;t are the conditional means of X t and Y t , respectively; hx;t and hy;t are the conditional variances of X t and Y t , respectively; and εt and ζ t are the two independent whitenoise processes with zero mean and unit variance. For the causality-in-mean test and the causality-in-variance test, we can use the following standardized innovation and the squares of the standardized innovation:
ut ¼
2.1. The LA-VAR technique
871
k ¼ 0; 7 1; 72; …
ð10Þ
k ¼ 0; 71; 72; …;
ð11Þ
and, similarly, where cεε ð0Þ, cζζ ð0Þ and cuu ð0Þ, cvv ð0Þ are defined as the sample variances of εt , ζ t and ut , vt , respectively. Causality in the mean and variance of X t and Y t can be tested by examining r^ εζ ðkÞ and r^ uv ðkÞ, respectively. Under the condition of regularity, the following holds: pffiffiffi L ð12Þ T r^ εζ ðki Þ-Nð0; 1Þ
872
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Fig. 1. Time series plots of the natural logarithms of Entergy, Henry Hub, and WTI.
Table 2 Summary statistics. Statistic
Entergy
Henry Hub
WTI
Observations Mean Median Maximum Minimum Standard deviation Skewness Kurtosis Jarque-Bera Q ð10Þ
1216 4.003 4.030 5.075 2.983 0.344 0.183 2.952 6.927 [0.031] 9548 [0.000] 9414 [0.000]
1216 1.927 1.956 2.733 0.920 0.342 0.240 3.080 11.961 [0.003] 11,523 [0.000] 11,453 [0.000]
1216 4.233 4.193 4.979 3.523 0.274 0.559 3.256 66.679 [0.000] 11,508 [0.000] 11,575 [0.000]
Q 2 ð10Þ
Note: The values in brackets indicate the p-values. Q ð10Þ is a test statistic for the null hypothesis; it indicates that no autocorrelation exists up to order 10. Q 2 ð10Þ is a test statistic for the null hypothesis; it indicates that no autocorrelation exists up to order 10 for squared value.
pffiffiffi L T r^ uv ðki Þ-Nð0; 1Þ
ð13Þ
L
where - shows the convergence in the distribution. These two statistics can be used to test the null hypothesis of no causality in mean and variance. To test for a causal relationship at a specified lag, k, we compare r^ εζ ðkÞ and r^ uv ðkÞ with the standard normal distribution. If the test statistic is larger than the critical value of the normal distribution, we reject the null hypothesis.
3. Data We use the wholesale electricity prices at Entergy in the Intercontinental Exchange (ICE), the natural gas prices at Henry Hub in the New York Mercantile Exchange (NYMEX), and the crude oil prices at WTI in the NYMEX. The electricity prices in the data are given in dollars per MWh, and are 24-h weighted average values of the closing future prices of the nearest delivery date when the ICE is open. The natural gas prices are the closing future prices of the earliest delivery date in dollars per MBTU. The crude oil prices, given in dollars per barrel, are the closing future prices of the nearest contract month. We use daily data from the period January 3, 2005–December 31, 2009. The data sets containing electricity prices, natural gas prices, and crude oil prices were obtained from the website of the United
States Department of Energy′s Energy Information Administration (http://www.eia.gov/). There are about 40 missing days for each variable. We use observations only from days when three variables are complete. It is reasonably believed that the influence of this assumption is small because there are more than 1200 observations for each variable. Fig. 1 provides the time plots of the natural logarithms of each variable, and Table 2 lists the summary statistics. Though each variable seems to influence each other on the face of this figure, we could not deny the spuriousness of the relationships among these prices without empirical analyses. Before we proceed with each causality test, we test for the stationarity status of all variables. Various kinds of unit root tests have been developed; however, it is not easy to say what unit root test technique is the most dominant from the viewpoint of power. Therefore, we adopt the augmented Dickey–Fuller (ADF) test which has a null hypothesis of a unit root. Appropriate lag orders are determined by minimizing the Schwarz information criterion (SIC). Moreover, we adopt the Kwiatkowski–Phillips–Schmidt– Shin (KPSS) test which has a null hypothesis of a stationarity. The results obtained through each unit root test applied to the natural logarithm of each variable are reported in Tables 3 and 4. That all variables have a unit root is acceptable. Moreover, we present the results of the unit root test for the first difference of each natural logarithm in Tables 3 and 4. The results tend to a
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
where ep, gp, and op are the natural logarithms of the electricity price, the natural gas price, and the crude oil price, respectively; t is the time trend; s is the lag index; k is the maximum lag length based on the SIC; α1 , α2 , αe , αg , αo , β1 , β2 , βe , βg , βo , γ 1 , γ 2 , γ e , γ g , and γ o are the coefficients; and εe , εg , and εo are the error terms assumed to be white noise and normally and identically distributed. As for the appropriate lag order, we select 2 based on the SIC.
rejection of the unit root hypothesis in all cases. Therefore, we may think that Entergy, Henry Hub, and WTI are I(1).
4. Empirical results 4.1. Causality tests using the LA-VAR technique 4.1.1. Estimation of the VAR model We estimate the vector autoregression (VAR) model as follows:
k
ept ¼ α1 þ α2 t þ ∑
ð14Þ
βe;s epts þ βg;s gpts þ βo;s opts þ εg;t ;
ð15Þ
ept ¼ α^ 1 þ α^ 2 t þ ∑
γ e;s epts þ γ g;s gpts þ γ o;s opts þ εo;t
ð16Þ
gpt ¼ β^ 1 þ β^ 2 t þ ∑
k
s¼1 k
opt ¼ γ 1 þ γ 2 t þ ∑
s¼1
4.1.2. Estimation of the LA-VAR model The test can be conducted with the following VAR model estimated by OLS:
αe;s epts þ αg;s gpts þ αo;s opts þ εe;t ;
s¼1
gpt ¼ β1 þ β2 t þ ∑
873
3
s¼1
3
s¼1
α^ e;s epts þ α^ g;s gpts þ α^ o;s opts þ ε^ e;t
ð17Þ
β^ e;s epts þ β^ g;s gpts þ β^ o;s opts þ ε^ g;t
ð18Þ
Table 3 The ADF unit root test results. Variable
Statistic
Natural logarithm of levels
First difference of natural logarithm
None
Constant
Constant+trend
None
Constant
Constant+trend
Entergy
No. of lags t-statistics Probabilities
4 0.15 0.63
4 3.34 0.01
4 4.06 0.01
3 21.97a 0.00
3 21.96a 0.00
3 21.96a 0.00
Henry Hub
No. of lags t-statistics Probabilities
0 0.36 0.55
0 1.91 0.33
0 2.48 0.34
0 37.20a 0.00
0 37.18a 0.00
0 37.17a 0.00
WTI
No. of lags t-statistics Probabilities
0 0.50 0.82
0 2.21 0.20
0 2.12 0.53
0 37.20a 0.00
0 37.20a 0.00
0 37.19a 0.00
Note: aIndicates that the unit root hypothesis is rejected at the 1% significance level.
Table 4 The KPSS unit root test results. Variable
Entergy Henry Hub WTI Critical value (1% level)
Statistic
LM-Statistic LM-Statistic LM-Statistic
Natural logarithm of levels
First difference of natural logarithm
Constant
Constant+trend
Constant
Constant+trend
1.45a 1.35a 0.85a 0.74
0.42a 0.37a 0.36a 0.22
0.10 0.09 0.14 0.74
0.06 0.06 0.08 0.22
Note: aIndicates that the stationary hypothesis is rejected at the 1% significance level.
Table 5 Causality test Wald statistics. Dependent variable
Statistics
Independent variables Entergy
Henry Hub
1851.58a 0.00
66.79a 0.00
0.46 0.79
1843.29a 0.00
0.24 0.89
Entergy
Wald statistics Probabilities
Henry Hub
Wald statistics Probabilities
0.05 0.97
WTI
Wald statistics Probabilities
3.89 0.14
Note: aIndicates that the null hypothesis of no Granger-causality is rejected at the 1% significance level.
2.90 0.23
WTI
1801.46a 0.00
874
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 6 Estimation results of AR-EGARCH model: xt ¼ cm þ ∑ki ¼ 1 ωi xti þ εt ; εt ¼
pffiffiffiffiffi ffiffiffiffiffiffi ffiffiffiffiffiffiffi þ ∑m ht ut ; Eðut Þ ¼ 0; E u2t ¼ 1;log ðht Þ ¼ cv þ ∑li ¼ 1 αi pεti þβ pεt1 i ¼ 1 γ i log ðhti Þ hti
ht1
Entergy
Henry Hub
WTI
Mean equation
cm ω1 ω2 ω3 ω4
0.001 (0.002) 0.046 (0.029) 0.137 (0.028)a 0.106 (0.028)a 0.098 (0.028)a
0.000 (0.001) 0.027 (0.028)
0.001 (0.001) –0.043 (0.028)
Variance equation
cv α β γ
0.409 (0.113)a 0.229 (0.042)a 0.002 (0.024) 0.954 (0.019)a
0.166 (0.061)a 0.120 (0.030)a 0.025 (0.017) 0.989 (0.007)a
0.154 (0.040)a 0.105 (0.022)a 0.057 (0.015)a 0.991 (0.004)a
Q ð15Þ
1.356 (0.070)a 19.331 [0.20] 10.284 [0.80]
1.352 (0.070)a 16.181 [0.37] 8.116 [0.92]
1.802 (0.104)a 8.796 [0.89] 22.671 [0.09]
GED parameter Diagnostic
Q 2 ð15Þ
Note: The numbers in parentheses are standard errors. The numbers in brackets are p-values. Q ð15Þ is a test statistic for the null hypothesis that there is no autocorrelation up to order 15 for standardized residuals. Q 2 ð15Þ is a test statistic for the null hypothesis that there is no autocorrelation up to order 15 for standardized residuals squared. a Indicates statistical significance at the 1% level.
3
opt ¼ γ^ 1 þ γ^ 2 t þ ∑
s¼1
γ^ e;s epts þ γ^ g;s gpts þ γ^ o;s opts þ ε^ o;t
ð19Þ
where circumflex (^)indicates an estimation by OLS. As explained in Section 2.1, the maximum order of integration (dmax ) must not exceed the true lag length of the VAR model (k ¼ 2) within the framework of the LA-VAR approach. The unit root test results illustrated in Tables 3 and 4 indicate that dmax ¼ 1. 4.1.3. Wald tests for causality The Wald test results are reported in Table 5. Statistically significant at the 1% level, we demonstrate that natural gas prices unidirectionally affect electricity prices in mean. However, except for this, we find no causality between electricity prices, natural gas prices, and crude oil prices. 4.2. Causality tests using the CCF approach 4.2.1. Estimation of univariate time-series models We estimate a series of univariate time-series models to allow for time variation in both the conditional mean and the conditional variance. We model the dynamics of the electricity price, the natural gas price, and the crude oil price using the AR(k)-EGARCH (l,m) process2. The EGARCH model developed by Nelson (1991) need not constrain the no-negative condition in the variance equation. The conditional mean and conditional variance are respectively expressed as follows: k
xt ¼ cm þ ∑ ωi xti þ εt ; εt ¼ i¼1
pffiffiffiffiffi ht ut ; Eðut Þ ¼ 0; E u2t ¼ 1;
ε l m ε ffiffiffiffiffiffiffiffi þβ pt1 ffiffiffiffiffiffiffiffiffi þ ∑ γ i log ðhti Þ: log ðht Þ ¼ cv þ ∑ αi pti hti ht1 i ¼ 1 i¼1
ð20Þ
ð21Þ
These models are applied to the differences in the electricity price, natural gas price, and crude oil price. Eqs. (20) and (21) are the expressions for the mean and variance, respectively. We assume that the error term ut has a generalized error distribution (GED). In this paper, each model is estimated by the method of maximum likelihood, and the lag lengths of the mean and variance 2 AR implies the autoregressive process, and EGARCH indicates exponential generalized autoregressive conditional heteroscedasticity.
Table 7 Cross-correlation analysis for the levels and squares of the standardized residuals: Electricity prices and natural gas prices. k
Test statistics Levels Entergy and Henry Hub ðkÞ
0 2.804a 1 7.501a 2 1.827 3 3.316a 4 2.741a 5 0.628 6 0.903 7 0.976 8 0.063 9 1.091 10 0.049
Squares Entergy and Henry Hub ðþkÞ
Entergy and Henry Hub ðkÞ
Entergy and Henry Hub ðþkÞ
0.133 0.126 0.530 0.073 1.231 1.778 1.451 0.931 1.255 1.011
2.884a 1.977 0.028 0.307 1.039 0.282 0.792 0.408 1.656 1.200 0.596
1.053 1.447 0.806 1.220 1.241 1.479 1.255 0.321 0.063 1.594
Note: aIndicates statistical significance at the 1% level.
equations are selected by SIC. In addition, the CCF approach needs no autocorrelation of the standardized residuals. Therefore, we check the specification of the models with the Ljung–Box test. Parameter estimates and their standard errors are reported in Table 6. The following models are thus selected: the AR(4)EGARCH(1,1) model for electricity prices, the AR(1)-EGARCH(1,1) model for natural gas prices, and the AR(1)-EGARCH(1,1) model for crude oil prices. As shown in Table 6, the coefficients of the GARCH term (γ) and the ARCH term (α) are statistically significant at the 1% level for electricity prices, natural gas prices, and crude oil prices. The GED parameter is also statistically significant for all variables. Each is estimated to be less than 2; therefore, the error terms are found to be fat tailed. The asymmetric parameter (β) is statistically significant for crude oil prices, but not for electricity and natural gas prices. The parameter of the AR term (ω) is statistically significant for neither natural gas prices nor crude oil prices. For electricity prices, the parameters of the AR term (ω2 , ω3 , and ω4 ) are statistically significant, but the parameters of the AR term (ω1 ) are not.
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 8 Cross-correlation analysis for the levels and squares of the standardized residuals: Electricity prices and crude oil prices. k
Test statistics Levels
Squares
Entergy and WTI Entergy and WTI Entergy and WTI Entergy and WTI ðkÞ ðþkÞ ðkÞ ðþkÞ 0 1 2 3 4 5 6 7 8 9 10
1.482 2.305 0.335 1.458 1.751 1.336 0.743 0.816 0.150 1.015 0.530
0.976 1.646 1.684 0.171 1.468 1.116 0.830 0.042 2.019 0.453
1.189 0.195 0.384 0.338 1.513 2.141 1.078 0.436 0.607 0.112 0.181
1.294 0.774 0.778 0.537 0.649 0.014 0.237 0.589 1.301 0.115
Table 9 Cross-correlation analysis for the levels and squares of the standardized residuals: Natural gas prices and crude oil prices. k
Test statistics Levels
0 1 2 3 4 5 6 7 8 9 10
Squares
Henry Hub and WTI ðkÞ
Henry Hub and WTI ðþkÞ
Henry Hub and WTI ðkÞ
Henry Hub and WTI ðþkÞ
13.952a 1.224 0.101 0.554 0.499 0.237 0.861 0.464 0.600 0.363 0.844
1.454 1.116 0.436 1.744 1.084 0.272 0.701 0.879 0.059 0.439
4.763a 0.108 0.279 1.029 1.029 0.321 0.499 0.781 0.743 1.799 2.382
0.279 0.129 1.109 1.852 0.617 1.946 0.049 0.425 1.029 0.453
Note: aIndicates statistical significance at the 1% level.
Table 6 also indicates the Ljung–Box test statistics at lag 15, Q ð15Þ and Q 2 ð15Þ, which are test statistics for the null hypothesis of no autocorrelation up to order 15 for standardized residuals and standardized residuals squared, respectively. Each null hypothesis is accepted for all variables.
4.2.2. Tests for causality in mean and causality in variance Using the Cheung and Ng (1996) procedure, we analyze the causality in mean and causality in variance on the basis of the empirical results of the AR-EGARCH model. We obtain evidence of causality in mean (variance) if one or more of the test statistics calculated from the (squares of the) standardized residuals are significantly different from 0. The test statistics have a standard normal distribution in large samples. The empirical results are reported in Tables 7–9. It is rational to assume that players in each market have certain responses to changes in market circumstances and the maximum reaction time is two weeks. Furthermore, many previous papers (for example Winnard et al. (1997), Bhar and Hamori (2006, 2008), Inagaki (2007), and Nakajima and Hamori (2012)) that adopted the CCF approach selected 5–12 lags to examine the Granger-causal
875
relationship. Therefore, we present the statistics until 10 lags. The “Levels” column gives the test statistic based on standardized residuals themselves. These are used for testing causality in mean. The test statistic in the “Squares” column, based on the squares of the standardized residuals, is used to test for causality in variance. Table 7 shows the results of the causality test between electricity and natural gas prices. We demonstrate that natural gas prices unidirectionally affect electricity prices in mean. Table 8 presents the results of the causality test between electricity and crude oil prices. Statistically significant at the 1% level, we find no causality in either mean or variance. As seen in Table 9, there is no causality in either mean or variance between natural gas and crude oil prices. The CCF approach can give us the same results as the LA-VAR approach.
5. Concluding remarks For this research, we performed tests of causality between wholesale electricity prices and natural gas and oil prices in the United States. For wholesale electricity prices, we used data from Entergy, a prominent trading hub in the southern states. For natural gas and oil prices, we used Henry Hub and WTI price data—the most widely used price references for natural gas and oil, respectively. To test for Granger-causality, we used the LA-VAR technique developed by Toda and Yamamoto (1995) and the CCF approach proposed by Cheung and Ng (1996). This research makes four contributions. First, it analyzes the causality between wholesale electricity prices and primary energy prices in the southern states. In our literature survey, we discovered research addressing the relationships between electricity prices and primary energy prices in northeastern and western wholesale markets but no similar research on southern-United States wholesale markets. Second, this research tests and confirms the causality hypothesis: Changes in natural gas prices are useful for predicting changes in electricity prices, which most prior research supports. Third, this work contributes to the body of findings on causality between electricity prices and primary energy prices, in which it is still difficult to establish a consensus, as there has been relatively little prior research on this issue. Finally, this is the first study to test causality in variance between wholesale electricity prices and primary energy prices in the United States. Despite the considerable amount of prior research analyzing causality in variance between assorted economic variables in a range of countries and regions, Nakajima and Hamori′s (2012) work on Japan′s wholesale electricity market, Muñoz and Dickey′s (2009) work on Spain′s wholesale electricity market and Serletis and Shahmoradi′s (2006) work on Canada′s wholesale electricity market are the only three so far to have studied causality in variance between wholesale electricity prices and primary energy prices. The results of this study are as follows: The LA-VAR and CCF test results both indicated causality in mean unilaterally from natural gas prices to wholesale electricity prices at the 1% significance level. These results are consistent with those of nearly all previous research. Therefore, we can conclude that this causality is very robust. Gas power plants play the role of supply capability for peak demand, so their fuel prices are thought to affect wholesale power market prices, which reflect changes in supply and demand moment by moment. It is reasonable to assume, as Emery and Liu (2002) pointed out, that the common practice by power companies of using natural gas as a marginal fuel for generating peak power is the primary explanation. Although Brown and Yücel (2008) did not find causality between the Henry Hub natural gas price and regional electricity prices,
876
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
they did present causality between regional natural gas prices and regional electricity prices. However, given that Entergy Hub and Henry Hub—the data for which were also examined in the present research—are the wholesale electricity trading hub and the natural gas trading hub, respectively, for the state of Louisiana—and that the latter is generally used as the national market price reference for natural gas, our test results may also reflect characteristics of regional gas prices. However, this study did not indicate the causal relationship in variance among gas prices and power prices, though Woo et al. (2006) argued, “the wholesale electricity market and the natural gas market are inextricably intertwined, because natural gas is a significant input for electricity generation.” We may not conclude the spill of the risk in the fuel market into the risk in the power market, while the fluctuation of the fuel price influences the fluctuation of the power price. These two approaches lead to no causality in variance or in mean between crude oil prices and power prices. To our knowledge, no other research has tested causality between oil prices and electricity prices; thus, we believe that additional work in this area would be highly relevant. Before these tests, we expected that the prices of crude oil would cause the prices of electricity since oil power plants also shoulder peaking supply. However, the results of this study did not back up our expectation. The peak power capability in Entergy′s service territory is 3756 MW, which consists of 84% gas/oil plants, 12% gas plants, 2% hydro plants, and 1% oil plants. Gas/oil plants chiefly burn natural gas; therefore, it is not wrong to expect that the fuel price of market participants′ peaking generators may impact this causality. The oil price hardly seems to be the cause of the wholesale power price since the output by oil power stations is too small in this area. We suppose that another possible reason that the prices of crude oil do not cause the prices of electricity is the feature of WTI futures. It is a well-known fact that WTI futures prices are determined in the global crude oil market dynamics. However, the demand and supply in Cushing, Oklahoma, and the savings in the United States have significant implications for the spot price, because the oil possesses a regional feature and the entire quantity is consumed domestically. The characteristics of this spot price are different from those of the Brent spot price, which reflects the conditions not only of Europe but also of Asia and Africa because Brent is exported. Therefore, we were interested in whether WTI futures price also has regional characteristics that give rise to causality only with regard to electricity trading hubs in southern states. However, we found no causality because the futures price appears to have the characteristics of a global index rather than a regional index. If we used the spot prices, we might not deny the causality between the prices of oil and power. We also found no causality in either mean or variance between WTI and Henry Hub prices. In the United States, the United Kingdom, Belgium, and the Netherlands, the price of natural gas is determined according to the hub approach, under which prices are based on demand and supply conditions, while in Japan and the European continent excluding Belgium and the Netherlands, the price is linked to a crude oil benchmark. These examinations were able to catch the above-mentioned relationship between the oil market and the gas market. If we examined the causal relationships between the crude oil price and the natural gas price in continental Europe or Japan, we might display causality. The CCF approach tests in this study did not detect the transmission of volatility from the peaking-generating fuel markets to the wholesale power market. The risk in the power market consists of supply-side risk and demand-side risk. The risk on the supply side is the risk of the power generation cost and powergenerating capacity. Fixed costs, such as depreciation, labor, and
repairs expense, will not become risk factors on a daily and/or hourly basis. Therefore, it is possible to disregard fixed-costs risk in interpreting this analysis result. Moreover, we need not consider the capacity risk, because the reliability of power plants is not extremely low. We should pay attention to the risk of variable cost—that is, fuel cost— in order to understand this causality in variance appropriately. These results of no volatility spillover mean that the supply-side risk does not influence the variance of the wholesale electric power price. It can be said that the power generation fuel portfolio and fuel procurement contracts are appropriate from the viewpoint of risk management. On the other hand, we found causality in mean from the gas prices to the electricity prices. Therefore, it is necessary to review the power plant portfolio in order to suppress power price when the natural gas price is thought to be a certain constant trend. Here we can conclude that the risk in the wholesale electricity market is largely due to demand-side risk. The hedge by the weather derivative is very effective because the daily and hourly demand fluctuation is mainly caused by the weather condition. The activation of the weather derivative market is important. Finally, we provide an economic summary of this study. The cost of power generation comprises a variable cost such as fuel procurement and fixed costs such as depreciation, labor, and repairs. Given this fact, we cannot deny that fuel price truly causes cost and that the fuel price Granger-causes the cost. Electricity is secondary energy obtained by the conversion of primary energy such as uranium, coal, natural gas, oil, and hydropower. The variable costs of nuclear power, hydropower, and coal-fired power plants are negligible, although huge capital investments are required. Therefore, in order to operate those plants as much as possible, the total installed capacity of nuclear power, hydropower, and coal-fired power plants never exceeds the load, which fluctuates depending on the season and/or time. On the other hand, oil and gas plants are used to handle the adjustments in hourly and daily changes in demand. Thus, the marginal generating cost is the procurement of oil and/or gas because power storage is technically difficult and very expensive. Some previous studies argue that the marginal fuel price Granger-causes the wholesale power price. However, the Granger-causality between fuel and power depends on the degree of integration of the two markets. Even if there is a true causality from fuel prices to power generating costs, there is no certainty of Granger-causality from fuel prices to power prices. Nakajima and Hamori (2012) demonstrate that there is no Granger-causality from international crude oil prices or the exchange rates to Japan′s wholesale electricity prices, not only in mean but also in variance. In other words, there is no integration of procurement and the sales markets from the point of view of both price and risk transmission. Moreover, Nakajima (2013) shows that even import fuel prices do not Granger-cause power prices in Japan, which depends on imports for most of its primary energy sources. In some cases, because of the low level of liquidity and the lack of skilled traders caused by inappropriate institutional design and/or inexperienced, the marginal generating cost has not affected power prices. This study does not detect Granger-causality in variance from natural gas prices to electricity prices, but suggests Grangercausality in mean. In other words, Entergy Hub has been integrated with Henry Hub not in terms of risk, but in terms of level. The Granger-causality in mean found in this study is consistent with other studies because Entergy is one of the oldest power markets and has a trading system equivalent to other power markets. This is the first study to examine the transmission of risk between wholesale electricity prices and primary energy prices in the United States. Therefore, the results should continue to serve as a benchmark. According to this study, we cannot expect the risk
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
of Entergy to be hedged with the derivative of Henry Hub and/or WTI. It is necessary that the derivatives of Entergy and/or the demand-side derivatives are substantial. In order to further develop this research, we propose three research themes. First, we should test Granger-causality among fuel, wholesale power, and retail power prices. This will help us understand the integration of fuel markets, wholesale power markets, and retail power markets. Moreover, it may reveal the efficiency of price formation. Second, we should apply non-linear methodology to the Granger-causality test, which has not been attempted in this study. The type of fuel affecting marginal generating cost depends on the power load, which tends to fluctuate. Therefore, if the load fluctuates widely, the relationship between fuel markets and the power market may be non-linear. A test using a non-linear model might detect Granger-causality, which this paper has not been able to detect. Finally, the addition of this research, which analyzes the wholesale electricity market in the southern region, to prior research efforts means that only the north-central region remains to be covered for a complete geographic view of the wholesale electricity market in the United States. We therefore look forward to a future study focusing on ComEd, Cinergy, the Midwest ISO, or other representative northcentral United States hubs to achieve complete geographic coverage. These research themes have great value in the future. References Amavilah, V.H.S., 1995. The capitalist world aggregate supply and demand model for natural uranium. Energy Economics 17, 211–220. Asche, F., Osmundsen, P., Sandsmark, M., 2006. The UK market for natural gas, oil and electricity: Are the prices decoupled? Energy Journal 27 (2), 27–40. Bhar, R., Hamori, S., 2006. Empirical investigation on the relationship between Japanese and Asian emerging equity markets. Applied Financial Economics Letters 2, 77–86.
877
Bhar, R., Hamori, S., 2008. Information content of commodity futures prices for monetary policy. Economic Modelling 25, 274–283. Brown, S.P.A., Yücel, M.K., 2008. Deliverability and regional pricing in U.S. natural gas markets. Energy Economics 30, 2441–2453. Cheung, Y.W., Ng, L.K., 1996. A causality-in-variance test and its application to financial market prices. Journal of Econometrics 72, 33–48. Emery, G.W., Liu, Q., 2002. An analysis of the relationship between electricity and natural-gas futures prices. Journal of Futures Markets 22, 95–122. Engle, R.F., Granger, C.W.J., 1987. Co-integration and error correction: representation, estimation and testing. Econometrica 55, 251–276. Granger, C.W.J., 1969. Investigating causal relations by econometric models and cross-spectral methods. Econometrica 37, 424–438. Inagaki, K., 2007. Testing for volatility spillover between the British pound and the euro. Research in International Business and Finance 21, 161–174. Mjelde, J.W., Bessler, D.A., 2009. Market integration among electricity markets and their major fuel source markets. Energy Economics 31, 482–491. Mohammadi, H., 2009. Electricity prices and fuels costs: Long-run relations and short-run dynamics. Energy Economics 31, 503–509. Nakajima, T., 2013. Inefficient and opaque price formation in the Japan Electric Power Exchange. Energy Policy 55, 329–334. Nakajima, T., Hamori, S., 2012. Causality-in-mean and causality-in-variance among electricity prices, crude oil prices, and yen-US dollar exchange rates in Japan. Research in International Business and Finance 26, 371–386. Muñoz, M.P., Dickey, D.A., 2009. Are electricity prices affected by the US dollar to Euro exchange rate? The Spanish case. Energy Economics 31, 857–866. Nelson, D.B., 1991. Conditional heteroskedasticity in asset returns: a new approach. Econometrica 59, 347–370. Serletis, A., Herbert, J., 1999. The message in North American energy prices. Energy Economics 21, 471–483. Serletis, A., Shahmoradi, A., 2006. Measuring and testing natural gas and electricity markets volatility: evidence from Alberta′s deregulated markets. Studies in Nonlinear Dynamics and Econometrics 10 (3), 1–20. Toda, H.Y., Yamamoto, Y., 1995. Statistical inference in vector autoregression with possibly integrated processes. Journal of Econometrics 66, 225–250. Winnard, P., Chang, F., Rusckowski, M., Mardirossian, G., Hnatowich, D.J., Hu, J.W., Chen, M., Fok, R.C.W., Huang, B., 1997. Causality in volatility and volatility spillover effects between US, Japan and four equity markets in the South China growth triangular. Journal of International Financial Markets, Institutions & Money 7, 351–367. Woo, C.K., Olson, A., Horowitz, I., Luk, S., 2006. Bi-directional causality in California′s electricity and natural-gas markets. Energy Policy 34, 2060–2070.
Contents lists available at ScienceDirect
Energy Policy journal homepage: www.elsevier.com/locate/enpol
Testing causal relationships between wholesale electricity prices and primary energy prices Tadahiro Nakajima a,n, Shigeyuki Hamori b,1 a b
The Kansai Electric Power Company, Incorporated, 3-6-16, Nakanoshima, Kita-Ku, Osaka 530-8270, Japan Faculty of Economics, Kobe University, 2-1, Rokkodai, Nada-Ku, Kobe 657-8501, Japan
H I G H L I G H T S
We test the Granger-causality among wholesale electricity and primary energy prices. We test not only the causality in mean but also the causality in variance. The results show that gas prices Granger-cause electricity prices in mean. We find no Granger-causality in variance among these variables.
art ic l e i nf o
a b s t r a c t
Article history: Received 31 July 2012 Accepted 8 July 2013 Available online 2 August 2013
We apply the lag-augmented vector autoregression technique to test the Granger-causal relationships among wholesale electricity prices, natural gas prices, and crude oil prices. In addition, by adopting a cross-correlation function approach, we test not only the causality in mean but also the causality in variance between the variables. The results of tests using both techniques show that gas prices Grangercause electricity prices in mean. We find no Granger-causality in variance among these variables. & 2013 Elsevier Ltd. All rights reserved.
Keywords: Electricity price Primary energy price Causality test
1. Introduction In the past, it was considered natural that competition was restricted and companies operated monopolistically in the electricity industry. Around 1990, however, the industry began to be liberalized, as many countries deregulated their electricity industries. The wholesale electricity industries in most states of the United States are now open to competition among players in the market. Market structures differ by geographical area, with negotiated- and organized-transaction markets being the two main market types. In a negotiated-transaction market, electricity transactions are directly negotiated by sellers and buyers, and individual transmission line owners establish their own electricity supply plans. On the other hand, in an organized-transaction market, an entity independent of market participants controls all transmission facilities and operates a spot market, with a few exceptions.
n Corresponding author. Tel.: +81 6 6441 8821; fax: +81 6 6441 3584. E-mail addresses: [email protected] (T. Nakajima), [email protected] (S. Hamori). 1 Tel./fax: +81 78 803 6832.
0301-4215/$ - see front matter & 2013 Elsevier Ltd. All rights reserved. http://dx.doi.org/10.1016/j.enpol.2013.07.033
In December 1999, the United States Federal Energy Regulatory Commission (FERC) issued Order No. 2000 regarding the formation of Regional Transmission Organizations (RTOs). This order asked all the electric utilities owning and operating a transmission system to participate in establishing an organization to operate the system over a large area and in a neutral manner. Entergy TransCo is one such RTO established on the basis of Order No. 2000. This type of RTO is a regulated, for-profit stock company that either owns or leases under long-term contracts all the transmission facilities within a specified area. It is the administrator and operator of the transmission system, responsible for investing in new transmission facilities. The Entergy wholesale market prices are determined at the regional trading hub. An empirical analysis of the relationship between wholesale electricity as a secondary energy source and crude oil, natural gas, coal, and uranium as primary energy sources would contribute greatly to the national energy policy and to the business and trading strategies of those in the energy business. Questions concerning the long-term relationship between the prices of primary energy sources and electricity—such as whether changes in international crude oil prices are useful in predicting regional wholesale electricity price movements or whether there is a volatility spillover effect between the wholesale electricity and
870
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 1 Causality tests of the relationships between electricity prices and primary energy prices in the United States. Sources
Study period Variables electricity Natural gas/Oil
Causality tests between Summary electricity prices and others
Brown and Yücel (2008)
1997–2007
PJM, PV
TZ6-PJM Topock-PV
Woo et al. (2006)
1999–2004
NP15, SP15 (CAISO) PG&E Citygate, SoCal
Emery and Liu (2002)
1996–2000
COB, PV
Henry Hub
PJM
Henry Hub, TZ6, No test among PJM Fuel oil for NY harbor and others
Serletis and Herbert (1999) 1996–1997
Henry Hub, TZ6, Topock
PG&E Citygate-NP15 SoCal-NP15 PG&E Citygate-SP15 SoCal-SP15 Henry Hub-COB Henry Hub-PV
There is bidirectional causality between regional gas and electricity prices, but no causality between gas wellhead and regional electricity prices. There is bidirectional causality between gas and electricity prices.
Electricity prices respond to departure from the equilibrium relationship, but gas prices do not. There appear to be effective arbitraging mechanisms for the prices of gas and fuel oil.
Note: “A-B” means that the past values of “A” give information about the future value of “B.” Serletis and Herbert (1999) argued that the electricity price series is I(0) and each of the other price series is I(1). Therefore, they inferred that the strength of the relationship between the power prices and each of the other prices would be spurious. PJM and PV represent wholesale electricity market prices in the eastern and western United States, respectively. Henry Hub, TZ6, and Topock represent a national natural gas market price, and natural gas market prices in the eastern and western United States, respectively. NP15 and SP15 represent wholesale electricity market prices in northern and southern California, respectively. PG&E Citygate and SoCal represent natural gas market prices in northern and southern California, respectively COB represents a wholesale electricity market price in northern California and Oregon. Fuel oil for NY harbor represents a standard reference price for oil in the U.S. Northeast.
primary energy markets—are of considerable interest from the viewpoint of energy security and energy company management. West Texas Intermediate (WTI) is a light sweet crude oil considered high-quality fuel for power generation. WTI, which is produced in Texas and is delivered at Cushing, Oklahoma, is consumed in the southern states, while the price of WTI is one of the most major benchmarks for the world oil market. Henry Hub is the pricing point for the natural gas in the pipeline system in Louisiana. The price at Henry Hub is seen to be the primary benchmark for the United States gas market. Recently, Henry Hub has started to be seen as the price index for the world natural gas market. Entergy′s service territory includes almost all of Louisiana, the southeastern part of Texas, the eastern three-quarters of Arkansas, and the western half of Mississippi. In other words, Entergy supplies electricity in the region that is a productive center and consuming area of oil and gas, whose prices are international and/or national benchmarks. This region has an extremely rare combination of features. This creates motivation to examine the southern energy market. By testing for the causal relationships among Entergy, WTI, and Henry Hub, we are able to confirm whether WTI and Henry Hub have features of the regional energy market as well as the international and/or national market. It is significant to investigate the causal relationship in the variance between the international and regional energy markets, and between primary and secondary energy markets, because of both electric power suppliers′ and electric power consumers′ concerns. Power suppliers desire to discuss the combination of various power sources, ideal fuel contracts, and the best oil stockpile in order to stabilize the cost of power procurement. Power consumers are interested in the selection and combination of energy sources of their facilities, equipment, and appliances in order to stabilize the total cost of energy procurement when they operate and make investments in these items. We apply the lag-augmented vector autoregression (LA-VAR) technique developed by Toda and Yamamoto (1995) to test for Granger-causality. The LA-VAR approach allows us to test for Granger-causality among time series variables in levels without detecting exactly their integration and cointegration properties.
In this respect, it is more advantageous to use the LA-VAR approach than the vector error correction model (VECM) approaches. The pretests for a unit root and cointegration in the economic time series and the estimation of the cointegrating vector are required before the causality test based on the VECM, and therefore the procedure is very complicated. Moreover, we adopt the cross-correlation function (CCF) approach developed by Cheung and Ng (1996) to test for Granger-causality in volatility as well as level. The traditional tests of causality, developed by Granger (1969) and Engle and Granger (1987), suffer from a number of problems:
The inability to address anything other than mean relationships. Model-building requirements. The need to pay attention to the omission of variables. The CCF approach, in contrast, offers the following advantages. It provides the ability to test for causality not only in the mean but also in the variance, and incorporates the use of univariate model residuals, which makes the building of a multivariate model unnecessary. Many studies empirically analyzed the Granger-causality between power prices and primary energy prices, because most economists can accept the economic model underlying such a Granger-causal relationship. We introduce some studies from the United States that have empirically analyzed the Granger-causality between power prices and primary energy prices using daily data. We summarize these in Table 1. Although researchers have analyzed the relationship between wholesale electricity prices and primary energy prices in the northeastern and western markets of the United States (for example PJM, PV, CAISO, and COB), no studies have examined the southern market (for example the Entergy Hub), to the best of our knowledge. It is difficult to find general or common causality relationships between electricity prices and the primary energy prices, because not many previous analyses exist. However, Brown and Yücel (2008), Woo et al. (2006), and Emery and Liu (2002) argue that past natural gas prices provide information about future wholesale electricity prices. In addition to these studies, Amavilah (1995) and Mohammadi (2009) tested the
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Granger-causality using annual data. Mjelde and Bessler (2009) investigated the Granger-causality using weekly data. Asche, et al. (2006), Serletis and Shahmoradi (2006), Muñoz and Dickey (2009), Nakajima and Hamori (2012), and Nakajima (2013) empirically examined the Granger-causality in the other countries. This paper makes four contributions to the literature. First, we test for a causal relationship between wholesale electricity prices and primary energy prices in southern states. To the best of our knowledge, no similar studies have examined the southern market. Second, we verify the causality hypothesis—which most previous studies propose—that the changes in natural gas prices are useful in predicting power price movements. Third, because few results of causality tests focus on the power market, we accumulate the analytical results. Finally, we analyze the causal relationship between United States wholesale power prices and primary energy prices in the variance for the first time, to our knowledge. The following section explains the LA-VAR technique and the CCF approach. Section 3 describes the data used. Section 4 presents the empirical testing results, and Section 5 summarizes our findings.
2. Methodology We apply the LA-VAR technique developed by Toda and Yamamoto (1995) and the CCF approach developed by Cheung and Ng (1996) in order to test for Granger-causality. The electricity, natural gas, and crude oil prices are probably subject to seasonality. However, it is not clear whether the relationship between these variables has seasonality. To our knowledge, no paper examines seasonality in the Granger-causal relationship between the power prices and other variables. Moreover, because we examine this relationship on the basis of daily data, the seasonality of each variable does not cause a bias in the results of our Granger-causality tests. Therefore, this study does not consider seasonality.
Toda and Yamamoto (1995) proposed a simple technique that requires prior tests of neither integration nor cointegration order. This technique allows the testing of coefficient restrictions in a level VAR model when the variables are of unknown integration or cointegration order. The method proceeds, briefly, as follows. Let the following equation generate fyt g, the n-dimensional vector constituting the level of the variables in the study: yt ¼ g o þ g 1 t þ J 1 yt1 þ J 2 yt2 þ … þ J k ytk þ εt ;
true values of J 1 , J 2 , …, J p are zero, those parameters are not included in restriction (2). Toda and Yamamoto (1995) establish that the Wald statistic asymptotically has a chi-square distribution with degrees of freedom equal to the number of excluded lagged variables regardless of the integration order of the process or the existence of a cointegrating relation. 2.2. The CCF approach Cheung and Ng (1996) developed a two-step procedure to test for causality not only in mean and but also in variance. The procedure is based on the residual CCF. The first step involves the estimation of univariate time-series models that allow for time variation in both conditional means and conditional variances. The second step is conducted by constructing the residuals and squared residuals, both standardized by conditional variances. The CCF of the standardized residuals is used to test the null hypothesis of no causality in mean, and the CCF of squaredstandardized residuals is used to test the null hypothesis of no causality in variance. Suppose two time series can be written as qffiffiffiffiffiffiffi X t ¼ μx;t þ hx;t εt ; ð4Þ Y t ¼ μy;t þ
t ¼ 1; 2; …; T
ð1Þ
where t is the time trend; k is the lag length; g o , g 1 , J 1 , J 2 , …, J k are the vectors or matrices of coefficients; and εt is an i.i.d. sequence of n-dimensional random vectors with zero mean and covariance matrix Σ ε . We formulate the null hypothesis—that the j-th variable does not Granger-cause the i-th variable—as follows: H 0 : J ij ð1Þ ¼ J ij ð2Þ ¼ … ¼ J ij ðkÞ ¼ 0;
yt ¼ γ^ 0 þ γ^ 1 t þ ^J 1 yt1 þ ^J 2 yt2 þ … þ ^J p ytp þ ε^ t
ð3Þ
where circumflex (^) indicates an estimation by OLS and p ¼ k þ dmax represents the true lag length k augmented by a suspected maximum integration order, dmax ðk≥dmax Þ. Considering that the
ð5Þ
Y t μy;t X t μx;t εt ¼ pffiffiffiffiffiffiffi ; ζ t ¼ pffiffiffiffiffiffiffi hy;t hx;t
ð6Þ
2 2 Y t μy;t X t μx;t ¼ ε2t ; vt ¼ ¼ ζ 2t : hx;t hy;t
ð7Þ
As both εt , ζ t and ut , vt are unobservable, we must use their estimates, ε^ t , ζ^ t and u^ t , v^ t to test the hypothesis of no causality in mean and variance, respectively. Next, we compute the sample cross-correlation coefficient at lag k, r^ εζ ðkÞ and r^ uv ðkÞ, from the consistent estimates of the conditional mean and variance of X t and Y t . This leaves us with cεζ ðkÞ r^ εζ ðkÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi cεε ð0Þcζζ ð0Þ
ð8Þ
cuv ðkÞ r^ uv ðkÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi cuu ð0Þcvv ð0Þ
ð9Þ
where cεζ ðkÞ and cuv ðkÞ are the k-th lag sample cross-covariance given by cεζ ðkÞ ¼
1 ^ ∑ð^εt ^ε Þðζ^ tk ζÞ; T
cuv ðkÞ ¼
1 ^ v^ tk v^ Þ; ∑ðu^ t uÞð T
ð2Þ
where J ij ðhÞ is the ði; jÞ element of the matrix J h ðh ¼ 1; 2; …; kÞ. The test can be conducted with the following VAR model, in level form, estimated by ordinary least squares (OLS):
qffiffiffiffiffiffiffi hy;t ζ t ;
where μx;t and μy;t are the conditional means of X t and Y t , respectively; hx;t and hy;t are the conditional variances of X t and Y t , respectively; and εt and ζ t are the two independent whitenoise processes with zero mean and unit variance. For the causality-in-mean test and the causality-in-variance test, we can use the following standardized innovation and the squares of the standardized innovation:
ut ¼
2.1. The LA-VAR technique
871
k ¼ 0; 7 1; 72; …
ð10Þ
k ¼ 0; 71; 72; …;
ð11Þ
and, similarly, where cεε ð0Þ, cζζ ð0Þ and cuu ð0Þ, cvv ð0Þ are defined as the sample variances of εt , ζ t and ut , vt , respectively. Causality in the mean and variance of X t and Y t can be tested by examining r^ εζ ðkÞ and r^ uv ðkÞ, respectively. Under the condition of regularity, the following holds: pffiffiffi L ð12Þ T r^ εζ ðki Þ-Nð0; 1Þ
872
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Fig. 1. Time series plots of the natural logarithms of Entergy, Henry Hub, and WTI.
Table 2 Summary statistics. Statistic
Entergy
Henry Hub
WTI
Observations Mean Median Maximum Minimum Standard deviation Skewness Kurtosis Jarque-Bera Q ð10Þ
1216 4.003 4.030 5.075 2.983 0.344 0.183 2.952 6.927 [0.031] 9548 [0.000] 9414 [0.000]
1216 1.927 1.956 2.733 0.920 0.342 0.240 3.080 11.961 [0.003] 11,523 [0.000] 11,453 [0.000]
1216 4.233 4.193 4.979 3.523 0.274 0.559 3.256 66.679 [0.000] 11,508 [0.000] 11,575 [0.000]
Q 2 ð10Þ
Note: The values in brackets indicate the p-values. Q ð10Þ is a test statistic for the null hypothesis; it indicates that no autocorrelation exists up to order 10. Q 2 ð10Þ is a test statistic for the null hypothesis; it indicates that no autocorrelation exists up to order 10 for squared value.
pffiffiffi L T r^ uv ðki Þ-Nð0; 1Þ
ð13Þ
L
where - shows the convergence in the distribution. These two statistics can be used to test the null hypothesis of no causality in mean and variance. To test for a causal relationship at a specified lag, k, we compare r^ εζ ðkÞ and r^ uv ðkÞ with the standard normal distribution. If the test statistic is larger than the critical value of the normal distribution, we reject the null hypothesis.
3. Data We use the wholesale electricity prices at Entergy in the Intercontinental Exchange (ICE), the natural gas prices at Henry Hub in the New York Mercantile Exchange (NYMEX), and the crude oil prices at WTI in the NYMEX. The electricity prices in the data are given in dollars per MWh, and are 24-h weighted average values of the closing future prices of the nearest delivery date when the ICE is open. The natural gas prices are the closing future prices of the earliest delivery date in dollars per MBTU. The crude oil prices, given in dollars per barrel, are the closing future prices of the nearest contract month. We use daily data from the period January 3, 2005–December 31, 2009. The data sets containing electricity prices, natural gas prices, and crude oil prices were obtained from the website of the United
States Department of Energy′s Energy Information Administration (http://www.eia.gov/). There are about 40 missing days for each variable. We use observations only from days when three variables are complete. It is reasonably believed that the influence of this assumption is small because there are more than 1200 observations for each variable. Fig. 1 provides the time plots of the natural logarithms of each variable, and Table 2 lists the summary statistics. Though each variable seems to influence each other on the face of this figure, we could not deny the spuriousness of the relationships among these prices without empirical analyses. Before we proceed with each causality test, we test for the stationarity status of all variables. Various kinds of unit root tests have been developed; however, it is not easy to say what unit root test technique is the most dominant from the viewpoint of power. Therefore, we adopt the augmented Dickey–Fuller (ADF) test which has a null hypothesis of a unit root. Appropriate lag orders are determined by minimizing the Schwarz information criterion (SIC). Moreover, we adopt the Kwiatkowski–Phillips–Schmidt– Shin (KPSS) test which has a null hypothesis of a stationarity. The results obtained through each unit root test applied to the natural logarithm of each variable are reported in Tables 3 and 4. That all variables have a unit root is acceptable. Moreover, we present the results of the unit root test for the first difference of each natural logarithm in Tables 3 and 4. The results tend to a
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
where ep, gp, and op are the natural logarithms of the electricity price, the natural gas price, and the crude oil price, respectively; t is the time trend; s is the lag index; k is the maximum lag length based on the SIC; α1 , α2 , αe , αg , αo , β1 , β2 , βe , βg , βo , γ 1 , γ 2 , γ e , γ g , and γ o are the coefficients; and εe , εg , and εo are the error terms assumed to be white noise and normally and identically distributed. As for the appropriate lag order, we select 2 based on the SIC.
rejection of the unit root hypothesis in all cases. Therefore, we may think that Entergy, Henry Hub, and WTI are I(1).
4. Empirical results 4.1. Causality tests using the LA-VAR technique 4.1.1. Estimation of the VAR model We estimate the vector autoregression (VAR) model as follows:
k
ept ¼ α1 þ α2 t þ ∑
ð14Þ
βe;s epts þ βg;s gpts þ βo;s opts þ εg;t ;
ð15Þ
ept ¼ α^ 1 þ α^ 2 t þ ∑
γ e;s epts þ γ g;s gpts þ γ o;s opts þ εo;t
ð16Þ
gpt ¼ β^ 1 þ β^ 2 t þ ∑
k
s¼1 k
opt ¼ γ 1 þ γ 2 t þ ∑
s¼1
4.1.2. Estimation of the LA-VAR model The test can be conducted with the following VAR model estimated by OLS:
αe;s epts þ αg;s gpts þ αo;s opts þ εe;t ;
s¼1
gpt ¼ β1 þ β2 t þ ∑
873
3
s¼1
3
s¼1
α^ e;s epts þ α^ g;s gpts þ α^ o;s opts þ ε^ e;t
ð17Þ
β^ e;s epts þ β^ g;s gpts þ β^ o;s opts þ ε^ g;t
ð18Þ
Table 3 The ADF unit root test results. Variable
Statistic
Natural logarithm of levels
First difference of natural logarithm
None
Constant
Constant+trend
None
Constant
Constant+trend
Entergy
No. of lags t-statistics Probabilities
4 0.15 0.63
4 3.34 0.01
4 4.06 0.01
3 21.97a 0.00
3 21.96a 0.00
3 21.96a 0.00
Henry Hub
No. of lags t-statistics Probabilities
0 0.36 0.55
0 1.91 0.33
0 2.48 0.34
0 37.20a 0.00
0 37.18a 0.00
0 37.17a 0.00
WTI
No. of lags t-statistics Probabilities
0 0.50 0.82
0 2.21 0.20
0 2.12 0.53
0 37.20a 0.00
0 37.20a 0.00
0 37.19a 0.00
Note: aIndicates that the unit root hypothesis is rejected at the 1% significance level.
Table 4 The KPSS unit root test results. Variable
Entergy Henry Hub WTI Critical value (1% level)
Statistic
LM-Statistic LM-Statistic LM-Statistic
Natural logarithm of levels
First difference of natural logarithm
Constant
Constant+trend
Constant
Constant+trend
1.45a 1.35a 0.85a 0.74
0.42a 0.37a 0.36a 0.22
0.10 0.09 0.14 0.74
0.06 0.06 0.08 0.22
Note: aIndicates that the stationary hypothesis is rejected at the 1% significance level.
Table 5 Causality test Wald statistics. Dependent variable
Statistics
Independent variables Entergy
Henry Hub
1851.58a 0.00
66.79a 0.00
0.46 0.79
1843.29a 0.00
0.24 0.89
Entergy
Wald statistics Probabilities
Henry Hub
Wald statistics Probabilities
0.05 0.97
WTI
Wald statistics Probabilities
3.89 0.14
Note: aIndicates that the null hypothesis of no Granger-causality is rejected at the 1% significance level.
2.90 0.23
WTI
1801.46a 0.00
874
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 6 Estimation results of AR-EGARCH model: xt ¼ cm þ ∑ki ¼ 1 ωi xti þ εt ; εt ¼
pffiffiffiffiffi ffiffiffiffiffiffi ffiffiffiffiffiffiffi þ ∑m ht ut ; Eðut Þ ¼ 0; E u2t ¼ 1;log ðht Þ ¼ cv þ ∑li ¼ 1 αi pεti þβ pεt1 i ¼ 1 γ i log ðhti Þ hti
ht1
Entergy
Henry Hub
WTI
Mean equation
cm ω1 ω2 ω3 ω4
0.001 (0.002) 0.046 (0.029) 0.137 (0.028)a 0.106 (0.028)a 0.098 (0.028)a
0.000 (0.001) 0.027 (0.028)
0.001 (0.001) –0.043 (0.028)
Variance equation
cv α β γ
0.409 (0.113)a 0.229 (0.042)a 0.002 (0.024) 0.954 (0.019)a
0.166 (0.061)a 0.120 (0.030)a 0.025 (0.017) 0.989 (0.007)a
0.154 (0.040)a 0.105 (0.022)a 0.057 (0.015)a 0.991 (0.004)a
Q ð15Þ
1.356 (0.070)a 19.331 [0.20] 10.284 [0.80]
1.352 (0.070)a 16.181 [0.37] 8.116 [0.92]
1.802 (0.104)a 8.796 [0.89] 22.671 [0.09]
GED parameter Diagnostic
Q 2 ð15Þ
Note: The numbers in parentheses are standard errors. The numbers in brackets are p-values. Q ð15Þ is a test statistic for the null hypothesis that there is no autocorrelation up to order 15 for standardized residuals. Q 2 ð15Þ is a test statistic for the null hypothesis that there is no autocorrelation up to order 15 for standardized residuals squared. a Indicates statistical significance at the 1% level.
3
opt ¼ γ^ 1 þ γ^ 2 t þ ∑
s¼1
γ^ e;s epts þ γ^ g;s gpts þ γ^ o;s opts þ ε^ o;t
ð19Þ
where circumflex (^)indicates an estimation by OLS. As explained in Section 2.1, the maximum order of integration (dmax ) must not exceed the true lag length of the VAR model (k ¼ 2) within the framework of the LA-VAR approach. The unit root test results illustrated in Tables 3 and 4 indicate that dmax ¼ 1. 4.1.3. Wald tests for causality The Wald test results are reported in Table 5. Statistically significant at the 1% level, we demonstrate that natural gas prices unidirectionally affect electricity prices in mean. However, except for this, we find no causality between electricity prices, natural gas prices, and crude oil prices. 4.2. Causality tests using the CCF approach 4.2.1. Estimation of univariate time-series models We estimate a series of univariate time-series models to allow for time variation in both the conditional mean and the conditional variance. We model the dynamics of the electricity price, the natural gas price, and the crude oil price using the AR(k)-EGARCH (l,m) process2. The EGARCH model developed by Nelson (1991) need not constrain the no-negative condition in the variance equation. The conditional mean and conditional variance are respectively expressed as follows: k
xt ¼ cm þ ∑ ωi xti þ εt ; εt ¼ i¼1
pffiffiffiffiffi ht ut ; Eðut Þ ¼ 0; E u2t ¼ 1;
ε l m ε ffiffiffiffiffiffiffiffi þβ pt1 ffiffiffiffiffiffiffiffiffi þ ∑ γ i log ðhti Þ: log ðht Þ ¼ cv þ ∑ αi pti hti ht1 i ¼ 1 i¼1
ð20Þ
ð21Þ
These models are applied to the differences in the electricity price, natural gas price, and crude oil price. Eqs. (20) and (21) are the expressions for the mean and variance, respectively. We assume that the error term ut has a generalized error distribution (GED). In this paper, each model is estimated by the method of maximum likelihood, and the lag lengths of the mean and variance 2 AR implies the autoregressive process, and EGARCH indicates exponential generalized autoregressive conditional heteroscedasticity.
Table 7 Cross-correlation analysis for the levels and squares of the standardized residuals: Electricity prices and natural gas prices. k
Test statistics Levels Entergy and Henry Hub ðkÞ
0 2.804a 1 7.501a 2 1.827 3 3.316a 4 2.741a 5 0.628 6 0.903 7 0.976 8 0.063 9 1.091 10 0.049
Squares Entergy and Henry Hub ðþkÞ
Entergy and Henry Hub ðkÞ
Entergy and Henry Hub ðþkÞ
0.133 0.126 0.530 0.073 1.231 1.778 1.451 0.931 1.255 1.011
2.884a 1.977 0.028 0.307 1.039 0.282 0.792 0.408 1.656 1.200 0.596
1.053 1.447 0.806 1.220 1.241 1.479 1.255 0.321 0.063 1.594
Note: aIndicates statistical significance at the 1% level.
equations are selected by SIC. In addition, the CCF approach needs no autocorrelation of the standardized residuals. Therefore, we check the specification of the models with the Ljung–Box test. Parameter estimates and their standard errors are reported in Table 6. The following models are thus selected: the AR(4)EGARCH(1,1) model for electricity prices, the AR(1)-EGARCH(1,1) model for natural gas prices, and the AR(1)-EGARCH(1,1) model for crude oil prices. As shown in Table 6, the coefficients of the GARCH term (γ) and the ARCH term (α) are statistically significant at the 1% level for electricity prices, natural gas prices, and crude oil prices. The GED parameter is also statistically significant for all variables. Each is estimated to be less than 2; therefore, the error terms are found to be fat tailed. The asymmetric parameter (β) is statistically significant for crude oil prices, but not for electricity and natural gas prices. The parameter of the AR term (ω) is statistically significant for neither natural gas prices nor crude oil prices. For electricity prices, the parameters of the AR term (ω2 , ω3 , and ω4 ) are statistically significant, but the parameters of the AR term (ω1 ) are not.
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
Table 8 Cross-correlation analysis for the levels and squares of the standardized residuals: Electricity prices and crude oil prices. k
Test statistics Levels
Squares
Entergy and WTI Entergy and WTI Entergy and WTI Entergy and WTI ðkÞ ðþkÞ ðkÞ ðþkÞ 0 1 2 3 4 5 6 7 8 9 10
1.482 2.305 0.335 1.458 1.751 1.336 0.743 0.816 0.150 1.015 0.530
0.976 1.646 1.684 0.171 1.468 1.116 0.830 0.042 2.019 0.453
1.189 0.195 0.384 0.338 1.513 2.141 1.078 0.436 0.607 0.112 0.181
1.294 0.774 0.778 0.537 0.649 0.014 0.237 0.589 1.301 0.115
Table 9 Cross-correlation analysis for the levels and squares of the standardized residuals: Natural gas prices and crude oil prices. k
Test statistics Levels
0 1 2 3 4 5 6 7 8 9 10
Squares
Henry Hub and WTI ðkÞ
Henry Hub and WTI ðþkÞ
Henry Hub and WTI ðkÞ
Henry Hub and WTI ðþkÞ
13.952a 1.224 0.101 0.554 0.499 0.237 0.861 0.464 0.600 0.363 0.844
1.454 1.116 0.436 1.744 1.084 0.272 0.701 0.879 0.059 0.439
4.763a 0.108 0.279 1.029 1.029 0.321 0.499 0.781 0.743 1.799 2.382
0.279 0.129 1.109 1.852 0.617 1.946 0.049 0.425 1.029 0.453
Note: aIndicates statistical significance at the 1% level.
Table 6 also indicates the Ljung–Box test statistics at lag 15, Q ð15Þ and Q 2 ð15Þ, which are test statistics for the null hypothesis of no autocorrelation up to order 15 for standardized residuals and standardized residuals squared, respectively. Each null hypothesis is accepted for all variables.
4.2.2. Tests for causality in mean and causality in variance Using the Cheung and Ng (1996) procedure, we analyze the causality in mean and causality in variance on the basis of the empirical results of the AR-EGARCH model. We obtain evidence of causality in mean (variance) if one or more of the test statistics calculated from the (squares of the) standardized residuals are significantly different from 0. The test statistics have a standard normal distribution in large samples. The empirical results are reported in Tables 7–9. It is rational to assume that players in each market have certain responses to changes in market circumstances and the maximum reaction time is two weeks. Furthermore, many previous papers (for example Winnard et al. (1997), Bhar and Hamori (2006, 2008), Inagaki (2007), and Nakajima and Hamori (2012)) that adopted the CCF approach selected 5–12 lags to examine the Granger-causal
875
relationship. Therefore, we present the statistics until 10 lags. The “Levels” column gives the test statistic based on standardized residuals themselves. These are used for testing causality in mean. The test statistic in the “Squares” column, based on the squares of the standardized residuals, is used to test for causality in variance. Table 7 shows the results of the causality test between electricity and natural gas prices. We demonstrate that natural gas prices unidirectionally affect electricity prices in mean. Table 8 presents the results of the causality test between electricity and crude oil prices. Statistically significant at the 1% level, we find no causality in either mean or variance. As seen in Table 9, there is no causality in either mean or variance between natural gas and crude oil prices. The CCF approach can give us the same results as the LA-VAR approach.
5. Concluding remarks For this research, we performed tests of causality between wholesale electricity prices and natural gas and oil prices in the United States. For wholesale electricity prices, we used data from Entergy, a prominent trading hub in the southern states. For natural gas and oil prices, we used Henry Hub and WTI price data—the most widely used price references for natural gas and oil, respectively. To test for Granger-causality, we used the LA-VAR technique developed by Toda and Yamamoto (1995) and the CCF approach proposed by Cheung and Ng (1996). This research makes four contributions. First, it analyzes the causality between wholesale electricity prices and primary energy prices in the southern states. In our literature survey, we discovered research addressing the relationships between electricity prices and primary energy prices in northeastern and western wholesale markets but no similar research on southern-United States wholesale markets. Second, this research tests and confirms the causality hypothesis: Changes in natural gas prices are useful for predicting changes in electricity prices, which most prior research supports. Third, this work contributes to the body of findings on causality between electricity prices and primary energy prices, in which it is still difficult to establish a consensus, as there has been relatively little prior research on this issue. Finally, this is the first study to test causality in variance between wholesale electricity prices and primary energy prices in the United States. Despite the considerable amount of prior research analyzing causality in variance between assorted economic variables in a range of countries and regions, Nakajima and Hamori′s (2012) work on Japan′s wholesale electricity market, Muñoz and Dickey′s (2009) work on Spain′s wholesale electricity market and Serletis and Shahmoradi′s (2006) work on Canada′s wholesale electricity market are the only three so far to have studied causality in variance between wholesale electricity prices and primary energy prices. The results of this study are as follows: The LA-VAR and CCF test results both indicated causality in mean unilaterally from natural gas prices to wholesale electricity prices at the 1% significance level. These results are consistent with those of nearly all previous research. Therefore, we can conclude that this causality is very robust. Gas power plants play the role of supply capability for peak demand, so their fuel prices are thought to affect wholesale power market prices, which reflect changes in supply and demand moment by moment. It is reasonable to assume, as Emery and Liu (2002) pointed out, that the common practice by power companies of using natural gas as a marginal fuel for generating peak power is the primary explanation. Although Brown and Yücel (2008) did not find causality between the Henry Hub natural gas price and regional electricity prices,
876
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
they did present causality between regional natural gas prices and regional electricity prices. However, given that Entergy Hub and Henry Hub—the data for which were also examined in the present research—are the wholesale electricity trading hub and the natural gas trading hub, respectively, for the state of Louisiana—and that the latter is generally used as the national market price reference for natural gas, our test results may also reflect characteristics of regional gas prices. However, this study did not indicate the causal relationship in variance among gas prices and power prices, though Woo et al. (2006) argued, “the wholesale electricity market and the natural gas market are inextricably intertwined, because natural gas is a significant input for electricity generation.” We may not conclude the spill of the risk in the fuel market into the risk in the power market, while the fluctuation of the fuel price influences the fluctuation of the power price. These two approaches lead to no causality in variance or in mean between crude oil prices and power prices. To our knowledge, no other research has tested causality between oil prices and electricity prices; thus, we believe that additional work in this area would be highly relevant. Before these tests, we expected that the prices of crude oil would cause the prices of electricity since oil power plants also shoulder peaking supply. However, the results of this study did not back up our expectation. The peak power capability in Entergy′s service territory is 3756 MW, which consists of 84% gas/oil plants, 12% gas plants, 2% hydro plants, and 1% oil plants. Gas/oil plants chiefly burn natural gas; therefore, it is not wrong to expect that the fuel price of market participants′ peaking generators may impact this causality. The oil price hardly seems to be the cause of the wholesale power price since the output by oil power stations is too small in this area. We suppose that another possible reason that the prices of crude oil do not cause the prices of electricity is the feature of WTI futures. It is a well-known fact that WTI futures prices are determined in the global crude oil market dynamics. However, the demand and supply in Cushing, Oklahoma, and the savings in the United States have significant implications for the spot price, because the oil possesses a regional feature and the entire quantity is consumed domestically. The characteristics of this spot price are different from those of the Brent spot price, which reflects the conditions not only of Europe but also of Asia and Africa because Brent is exported. Therefore, we were interested in whether WTI futures price also has regional characteristics that give rise to causality only with regard to electricity trading hubs in southern states. However, we found no causality because the futures price appears to have the characteristics of a global index rather than a regional index. If we used the spot prices, we might not deny the causality between the prices of oil and power. We also found no causality in either mean or variance between WTI and Henry Hub prices. In the United States, the United Kingdom, Belgium, and the Netherlands, the price of natural gas is determined according to the hub approach, under which prices are based on demand and supply conditions, while in Japan and the European continent excluding Belgium and the Netherlands, the price is linked to a crude oil benchmark. These examinations were able to catch the above-mentioned relationship between the oil market and the gas market. If we examined the causal relationships between the crude oil price and the natural gas price in continental Europe or Japan, we might display causality. The CCF approach tests in this study did not detect the transmission of volatility from the peaking-generating fuel markets to the wholesale power market. The risk in the power market consists of supply-side risk and demand-side risk. The risk on the supply side is the risk of the power generation cost and powergenerating capacity. Fixed costs, such as depreciation, labor, and
repairs expense, will not become risk factors on a daily and/or hourly basis. Therefore, it is possible to disregard fixed-costs risk in interpreting this analysis result. Moreover, we need not consider the capacity risk, because the reliability of power plants is not extremely low. We should pay attention to the risk of variable cost—that is, fuel cost— in order to understand this causality in variance appropriately. These results of no volatility spillover mean that the supply-side risk does not influence the variance of the wholesale electric power price. It can be said that the power generation fuel portfolio and fuel procurement contracts are appropriate from the viewpoint of risk management. On the other hand, we found causality in mean from the gas prices to the electricity prices. Therefore, it is necessary to review the power plant portfolio in order to suppress power price when the natural gas price is thought to be a certain constant trend. Here we can conclude that the risk in the wholesale electricity market is largely due to demand-side risk. The hedge by the weather derivative is very effective because the daily and hourly demand fluctuation is mainly caused by the weather condition. The activation of the weather derivative market is important. Finally, we provide an economic summary of this study. The cost of power generation comprises a variable cost such as fuel procurement and fixed costs such as depreciation, labor, and repairs. Given this fact, we cannot deny that fuel price truly causes cost and that the fuel price Granger-causes the cost. Electricity is secondary energy obtained by the conversion of primary energy such as uranium, coal, natural gas, oil, and hydropower. The variable costs of nuclear power, hydropower, and coal-fired power plants are negligible, although huge capital investments are required. Therefore, in order to operate those plants as much as possible, the total installed capacity of nuclear power, hydropower, and coal-fired power plants never exceeds the load, which fluctuates depending on the season and/or time. On the other hand, oil and gas plants are used to handle the adjustments in hourly and daily changes in demand. Thus, the marginal generating cost is the procurement of oil and/or gas because power storage is technically difficult and very expensive. Some previous studies argue that the marginal fuel price Granger-causes the wholesale power price. However, the Granger-causality between fuel and power depends on the degree of integration of the two markets. Even if there is a true causality from fuel prices to power generating costs, there is no certainty of Granger-causality from fuel prices to power prices. Nakajima and Hamori (2012) demonstrate that there is no Granger-causality from international crude oil prices or the exchange rates to Japan′s wholesale electricity prices, not only in mean but also in variance. In other words, there is no integration of procurement and the sales markets from the point of view of both price and risk transmission. Moreover, Nakajima (2013) shows that even import fuel prices do not Granger-cause power prices in Japan, which depends on imports for most of its primary energy sources. In some cases, because of the low level of liquidity and the lack of skilled traders caused by inappropriate institutional design and/or inexperienced, the marginal generating cost has not affected power prices. This study does not detect Granger-causality in variance from natural gas prices to electricity prices, but suggests Grangercausality in mean. In other words, Entergy Hub has been integrated with Henry Hub not in terms of risk, but in terms of level. The Granger-causality in mean found in this study is consistent with other studies because Entergy is one of the oldest power markets and has a trading system equivalent to other power markets. This is the first study to examine the transmission of risk between wholesale electricity prices and primary energy prices in the United States. Therefore, the results should continue to serve as a benchmark. According to this study, we cannot expect the risk
T. Nakajima, S. Hamori / Energy Policy 62 (2013) 869–877
of Entergy to be hedged with the derivative of Henry Hub and/or WTI. It is necessary that the derivatives of Entergy and/or the demand-side derivatives are substantial. In order to further develop this research, we propose three research themes. First, we should test Granger-causality among fuel, wholesale power, and retail power prices. This will help us understand the integration of fuel markets, wholesale power markets, and retail power markets. Moreover, it may reveal the efficiency of price formation. Second, we should apply non-linear methodology to the Granger-causality test, which has not been attempted in this study. The type of fuel affecting marginal generating cost depends on the power load, which tends to fluctuate. Therefore, if the load fluctuates widely, the relationship between fuel markets and the power market may be non-linear. A test using a non-linear model might detect Granger-causality, which this paper has not been able to detect. Finally, the addition of this research, which analyzes the wholesale electricity market in the southern region, to prior research efforts means that only the north-central region remains to be covered for a complete geographic view of the wholesale electricity market in the United States. We therefore look forward to a future study focusing on ComEd, Cinergy, the Midwest ISO, or other representative northcentral United States hubs to achieve complete geographic coverage. These research themes have great value in the future. References Amavilah, V.H.S., 1995. The capitalist world aggregate supply and demand model for natural uranium. Energy Economics 17, 211–220. Asche, F., Osmundsen, P., Sandsmark, M., 2006. The UK market for natural gas, oil and electricity: Are the prices decoupled? Energy Journal 27 (2), 27–40. Bhar, R., Hamori, S., 2006. Empirical investigation on the relationship between Japanese and Asian emerging equity markets. Applied Financial Economics Letters 2, 77–86.
877
Bhar, R., Hamori, S., 2008. Information content of commodity futures prices for monetary policy. Economic Modelling 25, 274–283. Brown, S.P.A., Yücel, M.K., 2008. Deliverability and regional pricing in U.S. natural gas markets. Energy Economics 30, 2441–2453. Cheung, Y.W., Ng, L.K., 1996. A causality-in-variance test and its application to financial market prices. Journal of Econometrics 72, 33–48. Emery, G.W., Liu, Q., 2002. An analysis of the relationship between electricity and natural-gas futures prices. Journal of Futures Markets 22, 95–122. Engle, R.F., Granger, C.W.J., 1987. Co-integration and error correction: representation, estimation and testing. Econometrica 55, 251–276. Granger, C.W.J., 1969. Investigating causal relations by econometric models and cross-spectral methods. Econometrica 37, 424–438. Inagaki, K., 2007. Testing for volatility spillover between the British pound and the euro. Research in International Business and Finance 21, 161–174. Mjelde, J.W., Bessler, D.A., 2009. Market integration among electricity markets and their major fuel source markets. Energy Economics 31, 482–491. Mohammadi, H., 2009. Electricity prices and fuels costs: Long-run relations and short-run dynamics. Energy Economics 31, 503–509. Nakajima, T., 2013. Inefficient and opaque price formation in the Japan Electric Power Exchange. Energy Policy 55, 329–334. Nakajima, T., Hamori, S., 2012. Causality-in-mean and causality-in-variance among electricity prices, crude oil prices, and yen-US dollar exchange rates in Japan. Research in International Business and Finance 26, 371–386. Muñoz, M.P., Dickey, D.A., 2009. Are electricity prices affected by the US dollar to Euro exchange rate? The Spanish case. Energy Economics 31, 857–866. Nelson, D.B., 1991. Conditional heteroskedasticity in asset returns: a new approach. Econometrica 59, 347–370. Serletis, A., Herbert, J., 1999. The message in North American energy prices. Energy Economics 21, 471–483. Serletis, A., Shahmoradi, A., 2006. Measuring and testing natural gas and electricity markets volatility: evidence from Alberta′s deregulated markets. Studies in Nonlinear Dynamics and Econometrics 10 (3), 1–20. Toda, H.Y., Yamamoto, Y., 1995. Statistical inference in vector autoregression with possibly integrated processes. Journal of Econometrics 66, 225–250. Winnard, P., Chang, F., Rusckowski, M., Mardirossian, G., Hnatowich, D.J., Hu, J.W., Chen, M., Fok, R.C.W., Huang, B., 1997. Causality in volatility and volatility spillover effects between US, Japan and four equity markets in the South China growth triangular. Journal of International Financial Markets, Institutions & Money 7, 351–367. Woo, C.K., Olson, A., Horowitz, I., Luk, S., 2006. Bi-directional causality in California′s electricity and natural-gas markets. Energy Policy 34, 2060–2070.